International Journal of Engineering Research and Applications (IJERA) is an open access online peer reviewed international journal that publishes research and review articles in the fields of Computer Science, Neural Networks, Electrical Engineering, Software Engineering, Information Technology, Mechanical Engineering, Chemical Engineering, Plastic Engineering, Food Technology, Textile Engineering, Nano Technology & science, Power Electronics, Electronics & Communication Engineering, Computational mathematics, Image processing, Civil Engineering, Structural Engineering, Environmental Engineering, VLSI Testing & Low Power VLSI Design etc.
Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
S0450598102
1. Saurabh D. Parmar et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 4, Issue 5( Version 5), May 2014, pp.98-102
www.ijera.com 98 | P a g e
Face Recognition in Various Illuminations
Saurabh D. Parmar, Vaishali J. Kalariya
Research Scholar, CE/IT Department-School of Engineering, R.K. University, Rajkot
Professor, CE/IT Department-School of Engineering, R.K. University, Rajkot
Abstract
Face Recognition (FR) under various illuminations is very challenging. Normalization technique is useful for
removing the dimness and shadow from the facial image which reduces the effect of illumination variations still
retaining the necessary information of the face. The robust local feature extractor which is the gray-scale
invariant texture called Local Binary Pattern (LBP) is helpful for feature extraction. K-Nearest Neighbor
classifier is utilized for the purpose of classification and to match the face images from the database.
Experimental results were based on Yale-B database with three different sub categories. The proposed method
has been tested to robust face recognition in various illumination conditions. Extensive experiment shows
that the proposed system can achieve very encouraging performance in various illumination environments.
Keywords: Face Recognition, Gamma Correction, Illumination, Dog Filtering, Image Preprocessing,
Equalization Of Normalization, Clahe (Contrast-Limited Adaptive Histogram Equalization), Log (Laplacian Of
Gaussian).
I. INTRODUCTION
Face recognition becomes an important task
in computer vision and one of the most successful
application areas recently. Face recognition
technologies have been motivated from the
application area of physical access, face image
surveillance, people activity awareness, visual
interaction for human computer interaction, and
humanized vision. The research on face recognition
has been conducted for more than thirty years, but,
still more processes and better techniques for facial
extraction and face recognition are needed. This
research work aims to resolve the classical problem
of human face recognition under dim light conditions
and develop an intelligent and efficient method for
recognizing human faces. The most existing
techniques like Eigen faces, Principal Component
Analysis, Elastic Graph Bunch Matching etc., has the
low dimension of the solution to classification and
generalization problems. Also, it is a challenging task
in a less-controlled environment like different
illumination variations for large databases. In order to
overcome the above limitations, an intelligent agent
is proposed to integrate three techniques viz.,
Illumination Normalization, Feature Extraction and
Classification for face recognition in various
illumination conditions to identify the match. This
research work focuses on the frontal view faces
subjecting under various illumination conditions. The
remaining paper is organized as follows: Section 2
describes some of the related works in face
recognition on various illumination conditions;
feature extraction techniques and the different
identification approaches. The Section 3 focuses on
recognizing the human face in various illumination
conditions using the proposed new scheme. In
Section 4, the proposed methods experimental and
their results are comprised. Section 5 outlines the
summary of the author’s conclusion.
II. RELATED WORK
There have been only a few studies reported
in the literature on utilizing various lighting
variations, which develops a historical and current
perspective of activities in the field of Face
Recognition.
Xiaoyang Tan et al [1] proposed
Illumination Normalization preprocessing chain
technique which incorporates a series of stages
designed to counter the effects of illumination
variations, local shadowing, and highlights while
preserving the essential elements of visual
appearance. The local texture feature sets called the
LBP (Local Binary Pattern)/LTP (Local Ternary
Pattern) descriptor used for the feature extraction and
improved KPCA classifier.
Padraig Cunnigham et al [5] highlighted that
k-NN process is transparent and so it is easy to
implement and it can be very effective if an analysis
of the neighbors is useful as explanation of the output
of the classifier is useful. Zeenathunisa et al [2]
implemented an integrated approach towards
recognizing the face under dim light conditions using
preprocessing chain, Local Binary Pattern and K-
Nearest Neighbor Classifier Zeenathunisa et al
[3]implemented approach towards recognizing face
in various dark illuminations they only consider
medium intensity level as a good face recognition.
Anila and Devarajan, et al [4] used preprocessing
RESEARCH ARTICLE OPEN ACCESS
2. Saurabh D. Parmar et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 4, Issue 5( Version 5), May 2014, pp.98-102
www.ijera.com 99 | P a g e
method is simple they concentrate for extended Yale-
b dataset. In their proposed preprocessing steps if
gamma correction has been applied to very low
illumination level image of edges of the face cannot
be identified.
III. CONCEPTUAL PREPROCESSING
STEPS FOR FACE
RECOGNITION
The proposed method combines the features
of CLAHE contrast enhancement, LOG filtering and
contrast equalization techniques. Over all stages of
proposed preprocessing method is shown in Fig 1.
The input image given to the image is preprocessed
to remove the dimness and local shades by using the
illumination normalization technique. Then the
features from the normalized image are extracted by
the agent using a gray-scale invariant texture measure
called Local Binary Pattern. Then the extracted
features are trained by the intelligent agent using the
classifier k-Nearest Neighbor to get the identified
image. Preprocessing stage of face recognition in
various illumination is given below fig.1
Fig.1 face recognition system steps
The face recognition in various illumination
systems some sequential steps to identify a matcher
from the existing data set. The procedure is as
follows:
ALGORITHM
Step 1: Accept the image as an input.
Step 2: The preprocessing phase follows the
following steps from 2.1 to 2.3 and repeat it for every
given image to remove the darkness and obtain the
brighter image.
Step 2.1: The gray-scale image is projected under
CLAHE which removes the darkness from the image
block by block window
Step 2.2: face-shaped edge from left lower jaw to
right lower jaw can be detected using LOG
Step 2.3: The final step of the preprocessor is to
apply equalization of normalization which rescales
the image intensity, still preserving the required
information of the input face image.
Step 3 : Apply Local Binary Pattern feature
descriptor for the preprocessed image for feature
extraction which stores the features in the form
of feature vector.
Step 4: These feature vectors are trained with the k–
NN recognizer to recognize the given image from the
training data set.
Step 5: If the recognized image exists in the training
set, then the recognizer identifies the image else does
not identified.
Step 6: Repeat the steps from 1 to 5 for each and
every input face image to find the matcher.
3.1 INPUT IMAGE
The Input Image for the preprocessor has
the dimension 159 x 159 which has been taken from
the Yale – B database. The images considered are
subjected to different light variations. One input face
image is projected to different illumination
variations.
3.2 PREPROCESSING
The proposed method combines the features
of CLAHE, LOG filtering and contrast equalization
techniques. Over all stages of proposed
preprocessing method is shown in Fig2
3.2.1 CLAHE (CONTRAST-LIMITED
ADAPTIVE HISTOGRAM EQUALIZATION)
Histogram equalization is to get an image
with uniformly distributed intensity levels over the
whole intensity scale. Histogram equalization might
produce the worse quality of result image than that of
the original image since the histogram of the result
image becomes approximately uniform. Large peaks
in the histogram can be caused by uninteresting area.
So, histogram equalization might lead to an increased
visibility of unwanted image noises. This means that
it does not adapt to local contrast requirement; minor
contrast differences can be entirely missed when the
number of pixels falling in a particular gray range is
relatively small An adaptive method to avoid this
drawback is block-based processing of histogram
equalization. In this method, image is divided into
sub-images or blocks, and histogram equalization is
performed to each sub-images or blocks.
Gray scale
Fig 2. Image before and after preprocessing sample
for proposed method
The CLAHE (Contrast-Limited Adaptive
Histogram Equalization) [12] introduced clip limit to
overcome the noise problem. The CLAHE limits the
amplification by clipping the histogram at a
3. Saurabh D. Parmar et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 4, Issue 5( Version 5), May 2014, pp.98-102
www.ijera.com 100 | P a g e
predefined value before computing the CDF. The
value at which the histogram is clipped, the so-called
clip limit, depends on the normalization of the
histogram and thereby on the size of the
neighborhood region. The redistribution will push
some bins over the clip limit again, resulting in an
effective clip limit that is larger than the prescribed
limit and the exact value of which depends on the
image.
In this experiment, size of neighborhood
region is set as fixed value as 8 x 8. This region is
actually a window. Some experiment has been made
and the window size gives good result and operates in
quite short of time. The bigger size of window, the
longer operation time consumed. Another parameter
that needs to be set up is clip-limit value. An adaptive
value of clip-limit has been set up in this experiment.
This value is the entropy of grayscale input image.
Image entropy becomes relatively low when
histogram is distributed on narrow intensity region
while image entropy becomes high when histogram is
uniformly distributed. Therefore, the entropy of the
histogram equalized image becomes higher than that
of the original input image. Entropy of image is
calculated by using [15]
𝐻 𝑥 = − 𝑝 𝑥𝑖 log2 𝑝(𝑥𝑖)𝑁
𝑖=1 Formula 3.1
As explained above that CLAHE works on a window
with certain size that adaptively enhances contrast
locally. It is different with gamma correction because
the adjustment of contrast and brightness is applied
directly to whole one image at one time.
3.2.2 LAPLACIAN OF GAUSSIAN (LOG)
Laplacian filters are derivative filters used to
find areas of rapid change (edges) in images. Since
derivative filters are very sensitive to noise, it is
common to smooth the image (e.g., using a Gaussian
filter) before applying the Laplacian. This two-step
process is call the Laplacian of Gaussian (LoG)
operation. Generally LoG is calculated by using
single equation below:
𝐿𝑜𝐺 = −
1
𝜋𝜎4 1 −
𝑥2+𝑦2
2𝜎2 𝑒
−
𝑥2+𝑦2
2𝜎2
Formula 3.2
In this experiment, we found that by using
sigma value = 1, gives better result in recognition.
Laplacian is a second derivative of image or first
derivative of Gaussian output. Laplacian is used as
high pass filter meanwhile, Gaussian as part of LoG
is used to smooth image to remove noise. Meanwhile,
DoG works as high pass filter but in different way.
Each Gaussian used has different value of Sigma.
After applying both Gaussian filters to same image,
then the difference is taken between both Gaussian
results. The subtracted result becomes the final result
of filtering. While using laplacian, we can get and
emphasize edges of object in image.
3.2.3 EQUALIZATION OF NORMALIZATION:
The final stage of the preprocessing chain
rescales the image intensities. It is important to
use a robust estimator because the signal typically
contains extreme values produced by highlights,
small dark regions such as nostrils, garbage at the
image borders, etc. One could use (for example)
the median of the absolute value of the signal for
this, but here a simpleand rapid approximation is
preferred based on a two stage process as follows:
I(x,y) ←
𝑖(𝑥,𝑦)
(𝑚𝑒𝑎𝑛 𝑥′,𝑦′ | 𝛼 )1
𝛼
Formula 3.3
I(x,y) ←
𝑖(𝑥,𝑦)
(𝑚𝑒𝑎𝑛 (min (𝜏,|𝐼 𝑥′,𝑦′ | 𝛼 )1
𝛼
Formula 3.4
Here, 𝛼 is a strongly compressive exponent that
reduces the influence of large values, t is a
threshold used to truncate large values after the
first phase of normalization, and the mean is over
the whole (unmasked part of the) image. By
default we use 𝛼 =0:1 t = 10[Tan and Triggs,
2010][6].
3.3 FEATURE EXTRACTOR
The well-known binaries features feature set
robust to illumination variations is the Local Binary
Patterns (LBP), which have been very effective for
face recognition tasks and gives a great performance
on standard face detection data sets [1]. Therefore,
the intelligent agent uses this elegant technique LBP
for feature extraction of its preprocessed images. In
the year 2006, Ahonen et al [9] introduced Local
Binary Pattern (LBP) features that the face image is
divided into several blocks or regions from which the
features are extracted and concatenated into an
enhanced feature vector. This approach has been
successfully utilized in face recognition and face
detection. The Local Binary Pattern is a highly
discriminative function and widely used in ordinary
analysis of facial appearances. The proposed
approach towards recognizing face under dim light
condition is very robust since the local binary
operator is subject to invariant against any monotonic
transformation of the gray scale. Another additive
advantage of this invariant operator is its
computational efficiency towards identifying the
features exactly. The LBP operator takes the local
neighborhood which is threshold at the gray level
value of the center pixel into a binary pattern. It is
defined for 3 x 3 neighborhood giving 8–bit integer
codes based on the eight pixels around the center
pixel.
LBP = (xx,yx) = 2n7
𝑛=0 s(in
_
𝑖 𝑐) Formula 3.5
In this case, where n runs over the 8
neighbors of the central pixel c, ic and in are
respectively the grey-level values of c and n,
and s(w) is 1 if w > 0 and 0 otherwise. LBP's are
4. Saurabh D. Parmar et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 4, Issue 5( Version 5), May 2014, pp.98-102
www.ijera.com 101 | P a g e
resistant to lighting effects in the sense that they
are invariant to monotonic gray-level
transformations, and they have been shown to
have high discriminative power for texture
classification [9]. However because they
threshold exactly at ic, they tend to be sensitive to
noise, es-specially in near-uniform image regions.
The LBP encoding process is illustrated in Fig.3
Fig.3 Illustration of the basic LBP operator
In [9], T. Ahonen et al. introduced an
LBP based method for face recognition that
divides the face into a regular grid of cells and
histograms the uniform LBP's within each cell,
finally using a nearest neighbor classifier over the
x2
histogram distance for recognition. Excellent
results were reported on the FERET dataset [1].
Some recent advances in using LBP for face
recognition are summarized in [2].
3.4 CLASSIFIER:
The classifier considered by the agent in this
research work is k-Nearest Neighbor (k-NN)
classifier which is a very intuitive method that
classifies unlabeled features based on their similarity
with features in their training sets. The k-NN
algorithm measures the distance between the query
face image and a set of images in the dataset. This is
a method for classifying objects based on closest
training examples in the feature space to determine
the person's identity or to make the
decision. It's a non-parametric lazy learning
algorithm which is applied to classification tasks.
Classification of nearest neighbor can be done by
comparing feature vectors of the different points of
the images. Using the weighted k-NN, where the
weights by which each of the k nearest points' class is
multiplied are proportional to the inverse of the
distance between that point and the point for which
the class is to be predicted also significantly
improves the results. Rabbani et al [13] described that
the simplest method for determining which face class
provides the best description of an input face image is
to find the face class k that minimizes the Euclidean
distance. The Euclidean distance is the straight line
distance between two pixels. Figure 6 represents the
common illustration of the image pixels and its
distance transform. the distance between the two
pixels can be computed using some distance function
d(x, y) where x, y are image pixels composed of N
features, such that x = {xh, x2,,...,xN} andy = {yh,
y2,...,yN}.The distance function considered here is the
Euclidean distance function:
𝑑 𝐸(x,y) = xt
2
− yt
2
𝑛
𝑡=1
Formula 3.
here the arithmetic mean across the dataset is
considered to be 0 and the standard deviation σ(x) to
be 1. Their corresponding functions can be described
as follows:
𝑥 =
1
𝑁
xt
𝑛
𝑡=1 Formula 3.7
σ (x)= (xt
𝑛
𝑡=1
− 𝑥)2
Formula 3.8
Image Distance Transform
Fig. 4 Illustration of Euclidean Distance
Using this distance measures which are used
to identify the alikeness of different items in the
database, the agent makes the classification. Once a
similarity measure is defined, each feature to be
classified will be compared to each predefined
feature which is extracted using LBP technique.
Whenever the classification is to be made for a new
input face image, its distance to each feature in the
training set must be determined. Only the k closest
entries in the training set are considered for matching.
Fig 5. Illustrates a process for k–NN classification
and the three closest points for a center pixel in the
training set are shown.
Fig 5. Classification of k-NN
IV. EXPERIMENTAL RESULT:
4.1 CALCULATING GAR AND FER:
In this experiment, we provide 150 images
that consist of 45 images of lowest dimness, 45
images of medium dimness and 45 images of highest
dimness. From 45 images of each subcategory, we
tried to combine number of images for training and
testing. The best result is when we use 6 images
which consist of 1 image from each subcategories of
each class. Class means varies of faces. In this
experiment, we use 15 different faces or class.
GAR is calculated by using formula below:
GAR (%) =
#𝐶𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 𝐼𝑚𝑎𝑔𝑒𝑠
#𝐼𝑚𝑎𝑔𝑒𝑠 𝑖𝑛 𝑇𝑟𝑎𝑖𝑛𝑖𝑛𝑔 𝐷𝑎𝑡𝑎𝑠𝑒𝑡
x 100
Formula 3.9
0
0
0
0 1 0
0 0 0
1.4
1
1.0 1.41
1.0 0.0 1.0
1.4
1
1.0 1.41
5. Saurabh D. Parmar et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 4, Issue 5( Version 5), May 2014, pp.98-102
www.ijera.com 102 | P a g e
The experiment was started by examining
face recognition method described in base paper for
organized images testing and training. From 45
images for testing in each subcategory, we observed
that recognition GAR for low is 95% with 2 false
recognized, high is 57% with 19 false recognized and
for medium is 93% with 3 false recognized. Fig 6.
Below depicts comparison between GAR and FER of
proposed method.
Fig.6 Proposed method Graph
V. Conclusion
A new technique of preprocessing has been
proposed for face recognition in various
illuminations. We got encouraging results. Also it
does not need much computation time. It can be
achieved by using a simple, efficient image
preprocessing phase that recognition performance
will be high when compared to the techniques where
face recognition is performed with typical
preprocessing. As a future vision, the most
elegant LBP technique can be utilized for multi–
view faces subjected under various illuminations for
face recognition.
REFERENCES
[1] Xiaoyang Tan and Bill Triggs, “Enhanced
Local Texture Feature Sets for Face
Recognition Under Difficult Lighting
Conditions”, IEEE Trans, on Image
Processing, Vol. 19, No.6, June 2010
[2] S. Zeenathunisa, A. Jaya and M.A. Rabbani,
“An Integrated Approach Towards
Recognizing Face Under Dim Light
Conditions”, In Proc. of 3rdNational Conf.
on Ad. Trends in Computing, pp. 21-
23,April 2011
[3] S. Zeenathunisa, A. Jaya and M.A.
Rabbani, “A BIOMETRIC APPROACH
TOWARDS RECOGNIZING FACE IN
VARIOUS DARK ILLUMINATIONS”, In
Proc. of 3rdNational Conf. on Ad. Trends in
Computing, pp. 21-23,April 2011
[4] S.Anila & Dr.N.Devarajan, “An efficient
Preprocessing Technique under difficult
Lighting Conditions,” in Proc. National
Conference on Emerging Trends in
Computer Communication and Informatics
(ETCCI-2011), March 10-11, 2011.
[5] Padraig Cunnigham and Sarah Jane
Dilany , “k – Nearest Neighbour
Classifier”, Technical Report UCD-CSI-
2007-4, March 2007
[6] Xiaoyang Tan and Bill Triggs,
“Enhanced Local Texture Feature Sets for
Face Recognition Under Difficult Lighting
Conditions”, IEEE Trans, on Image
Processing, Vol. 19, No.6, June 2010
[7] A.S. Georghiades, P.N. Belhumeur and D.J.
Kriegman “From Few to Many: Illumination
Cone Models for Face Recognition under
Differing Pose and Lighting,” IEEE TPAMI,
23(6): 643-660, 2001.
[8] R. Gross and V. Brajovic, “An image
preprocessing algorithm for illumination
invariant face recognition,” in Proc.
AVBPA, 2003, pp. 10-18.
[9] Timo Ahonen, Abdenour Hadid and Matti
Pietikäinen, “Face Description with Local
Binary Patterns: Application to Face
Recognition”,IEEE Trans. on Pattern
Analysis and Machine Intelligence, Vol 28,
No.
[10] pp. 2037 – 2041, Dec 2006. [10] .K. Lee, J.
Ho, and D. Kriegman, “Acquiring linear
subspaces for face recognition under
variable lighting,” IEEE Trans. Pattern
Analysis & Machine Intelligence, vol. 27,
no. 5, pp. 684–698, 2005.
[11] K. Messer, J.Matas, J. Kittler, J. Luettin and
G. Maitre, “XM2VTSDB: The Extended
M2VTS Database”, in Proc. International
Conference on Audio- and Video-Based
Person Authentication, pp.72-77, 1999.
[12] G. Aggarwal and R. Chellappa, “Face
Recognition in the presence of multiple
illumination sources”, In Proc. of the IEEE
International Conference on Computer
Vision (ICCV ‘05), pp. 1169 – 1176, 2005
[13] M.A. Rabbani and C. Chellappan, “An
Effective Approach to Frontal Face
Recognition Using Distance Measures”,In
the Proceedings of Asian Journal of
Information Technology, 4(12), pp. 1110 –
1115, 2005
[14] K. Zuiderveld: Contrast Limited Adaptive
Histogram Equalization, Academic Press
Inc. (1994)
[15] R. Gray: Entropy and Information Theory,
Springer-Verlag (1990)